arX

iv:m

ath/

0504

071v

1 [

mat

h.FA

] 5

Apr

200

5

Reproducing kernel Hilbert spaces and Mercer

theorem

C. Carmeli∗, E. De Vito†, A. Toigo‡,

1st February 2008

Abstract

We characterize the reproducing kernel Hilbert spaces whose ele-ments are p-integrable functions in terms of the boundedness of theintegral operator whose kernel is the reproducing kernel. Moreover,for p = 2 we show that the spectral decomposition of this integraloperator gives a complete description of the reproducing kernel.

1 Introduction

In recent years there is a new interest for the theory of reproducing ker-nel Hilbert spaces in different frameworks, like statistical learning theory(Cucker and Smale, 2002), signal analysis (Daubechies, 1992) and quantummechanics (Ali et al., 2000). In particular, for these applications there isoften the need of reproducing kernel Hilbert spaces having some additionalregularity property, as continuity or square-integrability.This paper is both a research article and a self-contained survey about thecharacterization of the reproducing kernel Hilbert spaces whose elements arecontinuous, measurable or p-integrable functions (1 ≤ p ≤ ∞). As brieflyreviewed in Section 2, this problem is equivalent to study the weak regularityproperties of maps taking values in an arbitrary Hilbert space (Saitoh, 1988).

∗C. Carmeli, Dipartimento di Fisica, Universita di Genova, and I.N.F.N., Sezione diGenova, Via Dodecaneso 33, 16146 Genova, Italy. e-mail: carmeli@ge.infn.it

†E. De Vito, Dipartimento di Matematica, Universita di Modena e Reggio Emilia, ViaCampi 213/B, 41100 Modena, Italy, and I.N.F.N., Sezione di Genova, Via Dodecaneso 33,16146 Genova, Italy. e-mail: devito@unimo.it

‡A. Toigo, Dipartimento di Fisica, Universita di Genova, and I.N.F.N., Sezione diGenova, Via Dodecaneso 33, 16146 Genova, Italy. e-mail: toigo@ge.infn.it

1

In the following sections, we take this last point of view and, given a set Xand a Hilbert space H, we characterize the weak regularity properties of amap γ : X → H in terms of the corresponding properties of the associatedkernel

X ×X ∋ (x, y) 7→ 〈γy, γx〉H ∈ C. (1)

More precisely, in Section 3 we prove that, if X is a measurable set and H is aseparable Hilbert space, γ is weakly measurable if and only if the associatedkernel (1) is separately measurable. Our proof is an easy consequence of theequivalence between weak and strong measurability. Moreover, if H is a spaceof square-integrable functions, we recall a result of Pettis (1938) clarifyingthe relation between vector valued maps and the theory of integral operators(Halmos and Sunder, 1978).In Section 4 we show that, if X is a measurable set endowed with a σ-finitemeasure µ and H is a separable Hilbert space, γ is weakly p-integrable ifand only if the integral operator with kernel given by Eq. (1) is bounded

from Lp

p−1 (X,µ) into Lp(X,µ) where 1 ≤ p ≤ ∞. Up to our knowledgethis result is new, though is deeply based on the theory of Pettis integral(Hille and Phillips, 1974). Moreover, for p = 1 we prove that the operator

H ∋ v 7→ 〈v, γ(·)〉H ∈ L1(X,µ)

is always compact. For finite measures this result is due to Pettis (1938), butit appears original for non finite measures. A brief discussion of the com-pactness for 1 < p < ∞ is also given to show that the strong p-integrabilityof the map γ is a sufficient condition, but is not necessary.In Section 5, if X is a locally compact space, it is proved that γ is weaklycontinuous if and only if the kernel (1) is locally bounded and separatelycontinuous, whereas the continuity of the kernel on X × X is equivalent tothe compactness of the operator

H ∋ v 7→ 〈v, γ(·)〉H ∈ C(X).

The results we present are due to Schwartz (1964) in the framework of repro-ducing kernel Hilbert spaces and we review them giving an elementary proofbased on standard functional analysis tools.Finally, in Section 6 if X is a locally compact second countable Hausdorffspace endowed with a positive Radon measure µ and H is a reproducing ker-nel Hilbert space such that H ⊂ L2(X,µ) ∩ C(X), we characterize the spaceH and the reproducing kernel Γ in terms of the spectral decomposition ofthe integral operator of kernel Γ. When X is compact, this kind of result isknown as Mercer theorem (Hochstadt, 1989). Extensions of Mercer theorem

2

can be found in Novitskii and Romanov (1999) and references therein. Ourproof is very simple and general since it is based on the polar decompositionof the canonical inclusion of H into L2(X,µ).

2 Notations

In this section we fix the notation, give the main definitions and reviewthe connection between reproducing kernel Hilbert spaces and vector valuedmaps.

If E is a (complex) Banach space, ‖·‖E denotes the norm of E, andE∗ is the Banach space of continuous antilinear functionals of E. We let〈·, ·〉 : E∗ × E → C be the canonical pairing. If F is another Banach spaceand A : E → F is a bounded linear operator, then A∗ : F ∗ → E∗ is theadjoint of A. If H is a Hilbert space, 〈·, ·〉H denotes the scalar product, linearin the first argument and, by means of the scalar product, H∗ is canonicallyidentified with H.

If X is a set, CX is the vector space of all the complex functions on X.If X is a measurable space endowed with a σ-finite positive measure µ,L0(X,µ) denotes the topological vector space of all measurable complex func-tions1 on X endowed with the topology of the convergence in measure onsubsets of finite measure (Schwartz, 1993). Given 1 ≤ p < ∞, Lp(X,µ) isthe Banach space of functions f ∈ L0(X,µ) such that ‖f‖p is µ-integrable,and L∞(X,µ) is the Banach space of elements f ∈ L0(X,µ) that are boundedµ-almost everywhere.

If X is a locally compact Hausdorff space, C(X) denotes the space ofcontinuous functions onX endowed with the open-compact topology (Kelley,1955). If X is second countable locally compact Hausdorff space, a positiveRadon measure on X is a positive measure µ defined on the Borel σ-algebraof X and finite on compact subsets, and supp µ denotes the support of µ.

Given a map γ from a set X into a Hilbert space H, we denote by Aγ :H → CX the linear operator

(Aγv)(x) = 〈v, γx〉H ∀x ∈ X v ∈ H, (2)

by Γ : X ×X → C the kernel

Γ(x, t) = 〈γt, γx〉H = Aγ(γt)(x) ∀x, t ∈ X (3)

and we letHγ = span γx ∈ H | x ∈ X, (4)

1As usual, a function is identified with its equivalence class µ-almost everywhere.

3

where span is the closure of the linear span.We now recall some basic definitions about weak regularity properties of

vector valued maps.

Definition 1 Let X be a set, H a Hilbert space and γ : X → H.

1. Assume X be a measurable space. The map γ is weakly measurable ifthe function Aγv is measurable for all v ∈ H.

2. Assume X be a measurable space endowed with a σ-finite measure µ

and 1 ≤ p ≤ ∞. The map γ is weakly p-integrable if the function Aγv

is p-integrable for all v ∈ H.

3. Assume that X is a locally compact Hausdorff space. The map γ isweakly continuous if the function Aγv is continuous for all v ∈ H.

The assumptions that µ is a σ-finite measure and X is locally compact willavoid technical problems. Moreover, the above properties clearly dependonly on Hγ. If X is a measurable space with a σ-finite measure µ and γ ismeasurable, we let

S = v ∈ H|µ(γ−1(B(v, ǫ))) > 0 ∀ǫ > 0 Hµ = spanS , (5)

where B(v, ǫ) = w ∈ H | ‖w − v‖H < ǫ. The closed subset S is the essentialrange of γ and S ⊂ Hγ so that Hµ ⊂ Hγ .

We need also the definition of kernel of positive type.

Definition 2 Given a set X, a complex kernel Γ : X ×X → C is called ofpositive type if

ℓ∑

i,j

cicjΓ(xi, xj) ≥ 0 (6)

for any ℓ ∈ N, x1, . . . , xℓ ∈ X and c1, . . . , cℓ ∈ C.

Remark 1 In the complex case, the positivity condition (6) ensures that

Γ(x, t) = Γ(t, x) ∀x, t ∈ X. (7)

This is no longer true in the real case. In this case, a kernel Γ : X ×X → R

is called of positive type if

Γ(x, t) = Γ(t, x) ∀x, t ∈ Xℓ

∑

i,j

cicjΓ(xi, xj) ≥ 0 ∀ℓ ∈ N, x1, . . . , xℓ ∈ X, c1, . . . , cℓ ∈ R.

4

In the framework of harmonic analysis, kernel of positive type are also calledpositive definite, and, in the context of reproducing kernel Hilbert spaces,Aronszajn kernel.

We now recall the definition of bounded kernel from the theory of integraloperators.

Definition 3 Let X and Y two measurable spaces endowed with σ-finitemeasures µ and ν, respectively. Fixed 1 ≤ p, r ≤ ∞, a measurable kernelK : X × Y → C is called (r, p)-bounded if, for all φ ∈ Lr(Y, ν),

1. there is a µ-null set Xφ ⊂ X such that the function K(x, ·)φ(·) is inL1(Y, ν) for all x 6∈ Xφ;

2. the map

x 7→∫

Y

K(x, y)φ(y) dν(y)

is in Lp(X,µ).

Moreover, if the following condition holds

1’) for µ-almost all x ∈ X, K(x, ·) ∈ L(Y, ν)r

r−1 (if r = 1, rr−1

= ∞ and,if r = ∞, r

r−1= 1),

K is called a Carleman (r, p)-bounded kernel.

Since ν is σ-finite, Condition 1’) is equivalent to Condition 1) and the factthat the µ-null set Xφ can be chosen in such a way to be independent of φ.In Halmos and Sunder (1978) there is an example of (2, 2)-bounded kernel,which is not a Carleman kernel. In the above definition, boundedness refersto the fact that the integral operator of kernel K is bounded from Lr(Y, ν)to Lp(X,µ), as shown in Proposition 4.

Finally, we recall that a reproducing kernel Hilbert space H on X is asubspace of CX such that H is a Hilbert space and, for all x ∈ X, there is afunction γx ∈ H satisfying

f(x) = 〈f, γx〉H ∀f ∈ H. (8)

The corresponding reproducing kernel of H is defined by

Γ(x, t) = 〈γt, γx〉H = γt(x) x, t ∈ X. (9)

We now review the connection between reproducing kernel Hilbert spaces,Hilbert space valued maps and kernels of positive type. The following resulthas been obtained by many authors, see Aronszajn (1950); Godement (1948);Kolmogorov (1992); Kreın (1949, 1950); Schoenberg (1938) and, for a com-plete list of references, Bekka and de la Harpe (2003); Hille (1972); Saitoh(1988, 1997); Schwartz (1964).

5

Proposition 1 Let X be a set.

1. Given a kernel Γ : X × X → C of positive type, there is a uniquereproducing kernel Hilbert space H on X with reproducing kernel Γ.The inclusion of H into CX is the operator Aγ associated with the mapγ : X → H defined by

γx = Γ(·, x) x ∈ X. (10)

The kernel (3) associated with γ is precisely the reproducing kernel ofH and

H = Hγ . (11)

2. Given a Hilbert space H and a map γ : X → H, the associated kernelΓ is of positive type and kerAγ = Hγ

⊥. In particular, Aγ is a uni-tary operator from Hγ onto the reproducing kernel Hilbert space withreproducing kernel Γ.

Proof. We report the proof given in Bekka and de la Harpe (2003).

1. Let γx = Γ(·, x) ∈ CX with x ∈ X and

H0 = span γx | x ∈ X ⊂ CX .

If f =∑

i ciγxi∈ H0 and g =

∑

j djγtj ∈ H0, the definition of γx

implies that

∑

ij

cidjΓ(tj , xi) =∑

j

djf(tj) =∑

i

cig(xi),

so the following sequilinear form on H0 ×H0

(f, g) 7→ 〈f, g〉 :=∑

ij

cidjΓ(tj, xi)

is well defined. Equation (7) ensures that 〈·, ·〉 is hermitian and Eq. (6)that 〈f, f〉 ≥ 0 for all f ∈ H0. Let now x ∈ X, the choice g = γx inthe above equation implies that

f(x) = 〈f, γx〉 ∀x ∈ X, (12)

for all f ∈ H0. Assume now that 〈f, f〉 = 0, the Cauchy-Schwarzinequality gives

|f(x)| ≤√

〈f, f〉√

〈γx, γx〉 = 0,

6

and, hence, f = 0. This shows that 〈·, ·〉 is a scalar product on H0.Finally, if (fn)n∈N is a Cauchy sequence in H0, Eq. (12) implies that, forall x ∈ X, (fn(x))n∈N converges to f(x) ∈ C, so that the completion ofH0 is a subspace of CX and Eq. (12) holds for all f ∈ H, showing thatH is a reproducing kernel Hilbert space with kernel Γ. The uniquenessis evident. Finally, let γ as in Eq. (10), Eq. (8) implies that the cor-responding operator Aγ is the inclusion of H into CX . Equation (11)follows by density of H0.

2. A simple check shows that the kernel Γ associated with γ is of positivetype and that kerAγ = Hγ

⊥. In particular, Aγ is a bijective map fromHγ onto Im Aγ. Let now HΓ be the reproducing kernel Hilbert spacewith kernel Γ given by point 1 of this proposition. Both Im Aγ andHΓ are subspaces of CX . In particular, for all x ∈ X, the functionΓ(·, x) = Aγγx is both in Im AΓ and in HΓ, see Eq. (2) and Eq. (9),respectively. Let now x, t ∈ X, the cited equations give

〈Aγγt, Aγγx〉HΓ= 〈Γ(·, t),Γ(·, x)〉HΓ

= Γ(x, t) = 〈γt, γx〉H,

showing thatAγ is an isometry from span γx | x ∈ X onto spanΓ(·, x) | x ∈X. The claim follows by definition of Hγ and Eq. (11) applied to HΓ.

The construction of the Hilbert space H given the kernel Γ is related to theGNS construction and it is also known as Kolmogorov theorem (Kolmogorov,1992). In the above proof, H is directly defined as a subspace of CX . Indoing so, one has to prove that the scalar product is well defined, but, dueto Eq. (8), it is strictly positive, so that there is no need to quotient withrespect to the vectors of null norm. Another possibility is to define H asa subspace of the formal linear combinations of elements Γ(·, xi), see, forexample, Dutkay (2004).

Proposition 1 shows that there is a one-to-one correspondence betweenHilbert space valued maps, complex kernels of positive type and reproduc-ing kernel Hilbert spaces. In particular, given a map γ, the Hilbert spacewhere the map takes value can be identified by means of the operator Aγ

with a unique reproducing kernel Hilbert space, which is a subspace of CX .Conversely, any reproducing kernel Hilbert space defines uniquely a Hilbertspace valued map γ such that the operator Aγ is the inclusion into C

X . Inboth cases, the weak regularity properties of γ stated in Definition 1 are incorrespondence with the regularity properties of the complex functions in thereproducing kernel Hilbert space. In the following, we will show that theseproperties can be completely characterized in terms of the kernel Γ. Due to

7

the above equivalence, we will state the results for an arbitrary Hilbert spacevalued map. Here, for the convenience of the reader, we summarize them inthe case of a separable reproducing kernel Hilbert space H and for X beinga locally compact second countable Hausdorff space endowed with a positiveRadon measure µ.

1. Proposition 2: H ⊂ L0(X,µ) if and only if Γ is separately measurable;

2. Proposition 5: H ⊂ Lp(X,µ) if and only if Γ is a ( pp−1

, p)-boundedkernel, for all 1 ≤ p ≤ ∞;

3. Proposition 8: H ⊂ C(X) if and only if Γ is separately continuous andlocally bounded.

The assumption that H is separable is essential (see Example 1) and, ingeneral, the inclusion of H into L0(X,µ) is not injective. However, if γ isweakly continuous and supp µ = X, H is separable and the inclusion isinjective (see Eqs. (11), (13), (22)).

3 Measurability

In this section we characterize the weak measurable maps and we discuss therelation with the theory of integral operators. The following proposition isbased on the well-known equivalence between weak and strong measurabilityfor maps with separable range (see, for example, Hille and Phillips (1974);Pettis (1938)).

Proposition 2 Let X be a measurable space, H a Hilbert space and γ : X →H. Assume that Hγ is separable, then the following conditions are equivalent:

1. the map γ is weakly measurable;

2. the map γ is measurable from X to H;

3. the function Γ is measurable from X ×X into C;

4. for all x ∈ X, the function Γ(x, ·) is measurable from X into C.

If X is endowed with a σ-finite measure µ and one of the above conditionsholds, then Im Aγ ⊂ L0(X,µ), the operator Aγ is a continuous from H intoL0(X,µ) and

ker Aγ = Hµ⊥ . (13)

Proof.

8

1) ⇒ 2) Since Hγ is separable, there is a denumerable Hilbert basis (en)n∈I ofHγ and

γx =∑

n∈I

(Aγen)(x) en ∀x ∈ X.

By definition Aγen = 〈en, γ(·)〉 and is measurable by assumption andγ is measurable since I is denumerable.

2) ⇒ 3) Since scalar product is continuous and Hγ is separable, the functionΓ(x, t) = 〈γt, γx〉H is measurable.

3) ⇒ 4) Given x ∈ X, the map t 7→ (x, t) is clearly measurable, so Condition (4)follows.

4) ⇒ 1) Let v =∑N

i=1 ciγxiwith xi ∈ X. By assumption Γ(·, xi) is measurable

for all i, so is Aγv =∑

i ciΓ(·, xi). By density and Eq. (4), Aγv ismeasurable for all v ∈ Hγ . If v ∈ Hγ

⊥, Aγv = 0 and, hence, ismeasurable.

Assume now one of the four equivalent conditions. Since pointwise conver-gence implies convergence in measure on subsets of finite measure (Schwartz,1993), Eq. (2) gives that Aγ is continuous from H to L0(X,µ).

We now prove ker Aγ = Hµ⊥. Since Hγ

⊥ ⊂ ker Aγ, without loss ofgenerality, we can assume that Hγ = H with H separable and, by definitionof Hµ, it is enough to show that ker Aγ = S⊥, where S is the essential rangeof γ.The set X0 = x ∈ X | γx 6∈ S is of null measure. Indeed, since S is closedand γ is (strongly) measurable, X0 is measurable. Moreover, by definition,if x ∈ X0, there is ǫx > 0 such that µ(γ−1(B(γx, ǫx))) = 0. Since Hγ isseparable, each member of the family B(γx, ǫx)x∈X0

is a union of membersof a fixed countable family of open sets Ωii∈N, with each Ωi contained inB(γx, ǫx) for some x ∈ X0. Clearly µ(γ−1(Ωi)) = 0 and X0 ⊂ ∪i∈Nγ

−1(Ωi),so that µ(X0) = 0.Let now v ∈ S⊥, then (Aγv)(x) = 0 for all x 6∈ X0 since γx ∈ S, butµ(X0) = 0, so that v ∈ ker Aγ . Conversely, let v ∈ ker Aγ. By contradiction,assume there is w0 ∈ S such that 〈v, w0〉 6= 0. Since the scalar product iscontinuous,

〈v, w〉 6= 0 ∀w ∈ B(w0, ǫ),

for some ǫ > 0. In particular, 〈v, γx〉 6= 0 for all x ∈ γ−1(B(w0, ǫ)). However,µ(γ−1(B(w0, ǫ))) > 0, since w0 ∈ S, so that Aγv 6= 0. Since this last factcontradicts the assumption, it follows that 〈v, w0〉 = 0 for all w0 ∈ S, thatis, v ∈ S⊥.

9

The following example (see Bourbaki (2004)) shows that the separability ofHγ is essential in the above proposition.

Example 1 Let X = [0, 1] with the Lebesgue measure and H be the Hilbertspace of functions v : [0, 1] → C such that

∑

x∈X

|v(x)|2 < +∞,

with scalar product

〈v, w〉H =∑

x∈X

v(x)w(x).

Let A be a non measurable subset of [0, 1] and γ

γx =

δx x ∈ A

0 x 6∈ A

where δx(y) = 0 if y 6= x and δx(x) = 1. Since (Aγv)(x) = 0 for all butdenumerable x, Aγv is measurable. However, the function

Γ(x, x) =

1 x ∈ A

0 x 6∈ A

is not measurable and, hence, Condition (2) and Condition (3) of Proposi-tion 2 can not be true.

We now clarify the relation with the theory of integral operators (for a com-plete account see Halmos and Sunder (1978), where weakly integrable mapsare called Carleman functions). The following proposition, due to Pettis(1938), characterizes the measurable maps taking value in L2(Y, ν) (see alsoTheorem 11.5 of Halmos and Sunder (1978)). For a converse result, seeProposition 7 below.

Proposition 3 Let X and Y two measurable sets endowed with two σ-finitemeasures µ and ν, respectively. Suppose L2(Y, ν) is separable. Given a weaklymeasurable map γ : X → L2(Y, ν), there exists a measurable function Kγ :X × Y 7−→ C and a µ-null set N ⊂ X such that, for x ∈ X \ N , one hasγx (y) = Kγ (x, y) for ν-almost all y ∈ Y .

The operator Aγ : L2(Y, ν) → L0(X,µ) associated with γ is the integraloperator with kernel Kγ and the kernel associated with γ is

Γ(x, x′) =

∫

Y

Kγ(x′, y)Kγ(x, y) dν(y),

for µ⊗ µ-almost all (x, x′) ∈ X ×X.

10

Proof. Let (Cn)n≥1 be an increasing sequence of measurable subsets of Xsuch that µ(Cn) < +∞ and X =

⋃

n Cn. For all n ∈ N, define by induction

A0 = ∅An =

x ∈ X | x ∈ Cn, x 6∈ An−1 and ‖γx‖L2(Y ) ≤ n,

Clearly, each An is a subset of finite measure, An ∩ Am = ∅ if n 6= m, and⋃

nAn = X. Define

γnx =

γx if x ∈ An

0 elsewhere.

If f ∈ L2 (X × Y, µ⊗ ν), by Fubini theorem f(x, ·) ∈ L2(Y, ν) for µ-almostall x ∈ X. Moreover, by separability of L2(Y, ν) the map x 7→ 〈f (x, ·) , γn

x 〉L2(Y )

is measurable. Let λn be the linear form on L2 (X × Y, µ⊗ ν) given by

λn (f) =

∫

X

〈f (x, ·) , γnx 〉L2(Y ) dµ (x) .

The linear form λn is bounded since

|λn (f)| ≤∫

X

∣

∣

∣〈f (x, ·) , γn

x 〉L2(Y )

∣

∣

∣dµ (x)

≤∫

X

‖f (x, ·)‖L2(Y ) ‖γnx‖L2(Y ) dµ (x)

≤[∫

X

‖f (x, ·)‖2L2(Y ) dµ (x)

]1/2 [∫

X

‖γnx‖2

L2(Y ) dµ (x)

]1/2

≤ n√

µ (An) ‖f‖L2(X×Y ) .

By Riesz lemma, there exists unique Λn ∈ L2 (X × Y, µ⊗ ν) such that

λn (f) =

∫

X×Y

f (x, y) Λn (x, y) dµ (x) dν (y) .

It is not restrictive to assume that Λn (x, ·) ∈ L2 (Y, ν) for all x. Let u ∈L2 (Y, ν). For every measurable subset C of X with µ(C) < +∞, choosingf (x, y) = χC (x) u (y), the above relations and Fubini theorem give

∫

C

〈u, γnx〉L2(Y ) dµ (x) =

∫

C

dµ (x)

∫

Y

u (y)Λn (x, y) dν (y)

=

∫

C

⟨

u,Λn(x, ·)⟩

L2(Y )dµ (x) ,

11

which implies 〈u, γnx〉L2(Y ) =

⟨

u,Λn(x, ·)⟩

L2(Y )for µ-almost all x. Letting u

vary over a denumerable dense subset of L2(Y, ν), we find that there existsa µ-null set Nn ⊂ X such that, for x ∈ X \Nn,

γnx (y) = Λn (x, y) (14)

for ν-almost all y ∈ Y . Hence the map

Kγ (x, y) =∑

n

χAn(x) Λn (x, y)

is finite, measurable and, for all x ∈ An\Nn,

Kγ (x, y) = γnx (y) = γx(y)

for ν-almost all y ∈ Y . Since X =⋃

nAn and N =⋃

nNn is a µ-null set, thefirst part of the proposition follows.The definition of Aγ implies that Aγ is the integral operator of kernel Kγ andthe last equation easily follows.

4 Integrability

In this section we characterize weakly p-integrable maps. First of all, werecall that the integral operator with a (r, p)-bounded kernel is continuous.

Proposition 4 With the assumptions of Definition 3, let K be a (r, p)-bounded kernel, then the operator LK : Lr(Y, ν) → Lp(X,µ)

(LKφ)(x) =

∫

Y

K(x, y)φ(y) dν(y) for µ−a.a. x ∈ X (15)

is bounded.

Proof. We report the proof of Halmos and Sunder (1978). Since LK isdefined on all the space Lr(Y, ν), it is enough to prove that it is closed. Let(φn)n∈N be a sequence such that it converges to φ in Lr(Y, ν) and ψn = LKφn

converges to ψ in Lp(X,µ). Possibly passing to a double subsequence, wecan assume that both (φn)n∈N and (ψn)n∈N converges almost everywhere andthat there is ω ∈ Lr(Y, ν) for which |φn(y)| ≤ ω(y) for ν-almost all y ∈ Y

(if r < ∞ it is a consequence of Fischer-Riesz theorem, see Lemma 3.9of Halmos and Sunder (1978), for r = +∞, it is trivial). Condition 1 of

12

Definition 3 and the fact that denumerable union of null sets is a null setimply that, for µ-almost all x ∈ X, the sequence (K(x, ·)φn(·))n∈N

convergesto K(x, ·)φ(·) for ν-almost all y ∈ Y and it is bounded almost everywhere by|K(x, ·)ω(·)| ∈ L1(Y, ν). The dominated convergence theorem gives that

limn→+∞

ψn(x) = limn→+∞

∫

Y

K(x, y)φn(y) dν(y) =

∫

Y

K(x, y)φ(y) dν(y)

for µ-almost all x ∈ X, so that ψ(x) =∫

YK(x, y)φ(y) dν(y) in Lp(X,µ).

We are now in position to state the main result of the paper.

Proposition 5 Let X be a measurable space, µ a σ-finite measure on X andHγ a separable Hilbert space. Let 1 ≤ p ≤ ∞, then the following conditionsare equivalent

1. the map γ is weakly p-integrable;

2. the kernel Γ is (q, p)-bounded with q = pp−1

.

If one of the above conditions holds, then Im Aγ ⊂ Lp(X,µ) and

(i) Aγ is a bounded linear operator from H into Lp(X,µ);

(ii) its adjoint A∗γ : Lq(X,µ) → H is given by

A∗γφ =

∫

X

φ(x)γx dµ(x), (16)

where φ ∈ Lq(X,µ) and the integral has to be understood in the weaksense (if p = ∞ and q = 1, in Eq. (16) A∗

γ is the restriction of theadjoint to L1(X,µ) ⊂ L∞(X,µ)∗);

(iii) Γ is a Carleman kernel and AγA∗γ = LΓ, where LΓ is the integral oper-

ator of kernel Γ given by Eq. (15).

Proof.

1) ⇒ 2) To show that Aγ is a bounded operator from H into Lp(X,µ) we followthe proof of Hille and Phillips (1974). Since γ is weakly p-integrable,Aγ is a linear operator from H to Lp(X,µ). We claim that it is closed.Indeed, let (vn)n∈N be a sequence that converges to v ∈ H and the

13

sequence (Aγvn)n∈N converges to φ ∈ Lp(X,µ). By construction, forµ-almost all x ∈ X,

limn→+∞

(Aγvn)(x) = limn→+∞

〈vn, γx〉H = 〈v, γx〉H = (Aγv)(x).

By uniqueness of the limit, Aγv = φ, so that Aγ is closed. The closedgraph theorem implies that Aγ is bounded.We show Eq. (16). Given φ ∈ Lq(X,µ) and v ∈ H, by assump-tion Aγv ∈ Lp(X,µ), so that the function φAγv = φ(·)〈γ(·), v〉H isin L1(X,µ). Since v is arbitrary, it follows that the function φ(·) γ(·)is weakly integrable and

∫

X

φ(x) 〈γx, v〉H dµ(x) = 〈φ,Aγv〉 =⟨

A∗γφ, v

⟩

H ,

so that Eq. (16) holds.The fact that Γ is a Carleman kernel follows observing that, for all x ∈X, Γ(·, x) = Aγ(γx)(·), which is in Lp(X,µ) by assumption. Moreover,given φ ∈ Lq(X,µ), Eqs. (3) and (16) imply that∫

X

Γ(x, t)φ(t) dµ(t) =⟨

A∗γφ, γx

⟩

H = (AγA∗γφ)(x) for µ−a.a. x ∈ X.

Hence,∫

XΓ(·, t)φ(t) dµ(t) is in Lp(X,µ), that is, Γ is a Carleman (q, p)-

bounded kernel, and AγA∗γ = LΓ.

2) ⇒ 1) Since Γ is measurable and Hγ is separable, Proposition 2 ensures thefunction 〈v, γ(·)〉H be measurable for all v ∈ H. So it is enough toprove that |〈v, γ(·)〉H| ∈ Lp(X,µ)To this aim, let (An)n∈N be the sequence of sets defined in the proofof Proposition 3. For all n ∈ N let Dn =

⋃nk=1Ak, which is of finite

measure and γ is bounded by n on Dn.Let now φ ∈ Lq(X,µ), then the map x 7→ χDn

(x)φ(x)γx is clearlymeasurable and bounded by χDn

n|φ| ∈ L1(X,µ) since Dn has finitemeasure.Hence the linear operator Bn : Lq(X,µ) → H

Bnφ =

∫

X

χDn(x)φ(x)γx dµ(x) φ ∈ Lq(X,µ),

where the integral is in the strong sense, is well defined and boundedby

‖Bnφ‖H ≤ n

∫

Dn

|φ(x)| dµ(x) ≤ nµ(Dn)1/p‖φ‖q.

14

Since Γ is a bounded kernel, LΓ is a bounded operator and the followinginequality holds

‖Bnφ‖H2 =

∫

X

χDn(y)φ(y)

(∫

X

χDn(x)φ(x) 〈γx, γy〉H dµ(x)

)

dµ(y)

= 〈LΓ(χDnφ), (χDn

φ)〉 ≤ ‖LΓ‖‖φ‖q2,

for all φ ∈ Lq(X,µ), so that

supn∈N

‖Bn‖ ≤ M,

with M =√

‖LK‖ and, hence, supn ‖B∗n‖ ≤M .

Let now v ∈ H and φ ∈ Lq(X,µ), then

〈φ,B∗nv〉 = 〈Bnφ, v〉H =

∫

X

φ(x)〈χDn(x)γx, v〉H dµ(x) ∀φ ∈ Lq(X,µ),

so B∗nv(x) = χDn

(x)〈v, γx〉H for µ-almost all x ∈ X (in particular,B∗

nv ∈ L1(X,µ) if q = ∞), and

‖B∗nv‖p ≤ ‖B∗

n‖‖v‖H ≤M‖v‖H. (17)

Since⋃

n∈NDn = X, limn→+∞B∗

nv(x) = 〈v, γx〉H µ-almost everywhere,and Eq. (17) immediately ensures that 〈v, γ(·)〉H ∈ L∞(X,µ) for p =∞. If 1 ≤ p < ∞, the monotone convergence theorem gives that themap x 7→ |〈v, γx〉H|p is in L1(X,µ).

If γ is a weakly measurable function such that∫

X×X

|Γ(x, t)|p d(µ⊗ µ)(x, t) < +∞,

Fubini theorem and Holder inequality imply that Γ is a Carleman (q, p)-bounded kernel and, hence, γ is a weakly p-integrable function.The following proposition discusses the problem of compactness, comparewith Corollary 2. For finite measures the first statement is due to Pettis(1938). The second statement is well known. See, for example, Halmos and Sunder(1978) for a complete discussion about the compactness of integral operatorsin L2(X,µ).

Proposition 6 With the notation of Proposition 5, if Hγ is separable and γis weakly integrable, then Aγ is a compact operator in L1(X,µ).If p < ∞ and γ is strongly p-integrable, then Aγ is a compact operator inLp(X,µ).

15

Proof. Assuming that Hγ is separable and γ is weakly integrable, we provethat Aγ is compact. First, suppose that γ takes only a countable number ofvalues, that is,

γ =∑

n∈I

χEnvn,

where I is denumerable, the sequence (En)n∈I is disjoint, each En is of finitemeasure and vn ∈ H. The condition that γ is weakly integrable implies that

‖Aγv‖L1(X) =∑

n

µ(En)| 〈v, vn〉 | = ‖Tv‖ℓ1,

where T : H → ℓ1, (Tv)n = µ(En) 〈v, vn〉. Since T is a bounded operator fromthe Hilbert space H to ℓ1, it is known that T is compact (see Conway (1990)).Hence Aγ maps weakly convergent sequences into strongly convergent ones,so that it is compact.Assume now that γ is arbitrary, we claim there exists γ1 and γ2 such thatγ1 takes only a countable number of values, γ2 is strongly integrable andγ = γ1 + γ2. To this aim, let (An)n∈N be the sequence of sets defined inthe proof of Proposition 3. Given n ∈ N, since An has finite measure andγ is bounded by n on it, χAn

γ is strongly integrable, so there is a mapηn : An → H, which takes only a finite number of values and

∫

An

‖γ(x) − ηn(x)‖H dµ(x) ≤ 1

2n.

Since the sequence (An)n∈N is disjoint and measurable, the map

γ1 =∑

n∈N

χAnηn

is well defined, measurable and takes only a countable number of values. Letγ2 = γ − γ1, then

∫

X

‖γ2(x)‖H dµ(x) =∑

n

∫

An

‖γ(x) − ηn(x)‖H dµ(x) ≤∑

n

1

2n= 2,

so that γ2 is strongly integrable and γ = γ1 + γ2.Finally, since γ and γ2 are weakly integrable, so is γ1 and, by the previousresult, Aγ1

is compact. Since γ2 is strongly integrable, it is easy to check thatAγ2

is compact (see the second part of this proposition). The thesis followssince Aγ = Aγ1

+ Aγ2.

The second statement of the proposition is standard. Indeed, if γ is a stronglyp-integrable map, that is, γ is measurable and

∫

X

‖γx‖Hp dµ(x) =

∫

X

Γ(x, x)p/2 dµ(x) < +∞,

16

then Aγ : H → Lp(X,µ) is a compact operator. Indeed, if (vn)n∈N is asequence in H which converges weakly to 0, then 〈vn, γx〉H → 0 for all x ∈X. Since |〈vn, γx〉H| ≤ ‖vn‖H‖γx‖H and supn ‖vn‖H < +∞, it follows bydominated convergence theorem that Aγvn → 0 in Lp(X,µ). This showsthat Aγ maps weakly convergent sequences into norm convergent sequences,so Aγ is compact.

Example 2 below shows that, if p > 1 and γ is only weakly p-integrable,Aγ can be a non-compact operator in Lp(X,µ). However, if µ(X) < +∞,since Lp(X,µ) ⊂ L1(X,µ), Aγ is a compact operator from H into L1(X,µ)(Halmos and Sunder, 1978).If p = ∞ and γ is essentially bounded, Aγ is not necessarily compact.

For p = 2 we can compute A∗γAγ , which is known as frame operator in

the context of frame theory (see, for example, Young (2001)). The followingresult can be found in Halmos and Sunder (1978).

Corollary 1 With the notation of Proposition 5, assume that γ is weaklysquare-integrable, then

A∗γAγ =

∫

X

〈·, γx〉Hγx dµ(x), (18)

where the integral converges in the weak operator topology.In particular, Aγ is a Hilbert-Schmidt operator if and only if γ is strongly

square-integrable, that is,

∫

X

Γ(x, x) dµ(x) < +∞. (19)

If this last condition holds, the integral in Equation (16) converges in normand the integral in Equation (18) converges in trace norm.

Finally, γ is a (strongly) square-integrable function if and only if LΓ is atrace class operator from L2(X,µ) in L2(X,µ).

Proof. Eq. (18) follows from Eq. (16) and the definition of Aγ . To proveCondition (19), let (en)n∈N be a Hilbert basis of Hγ . Since A∗

γAγ is a positiveoperator and |〈γ(·), en〉H|2 is a positive function, by monotone convergence

17

theorem, one has that

Tr (A∗γAγ) =

∑

n

∫

X

|〈en, γx〉H|2 dµ(x)

=

∫

X

∑

n

|〈en, γx〉H|2 dµ(x)

=

∫

X

〈γx, γx〉H dµ(x)

=

∫

X

Γ(x, x) dµ(x)

and the thesis follows. Finally, we prove the statements about Eqs. (16)and (18). Indeed, since Hγ is separable, by Proposition 2, the map γ ismeasurable, so that the maps x 7→ φ(x)γx and x 7→ 〈·, γx〉Hγx are measurableas maps taking values in Hγ and in the separable Banach space of trace classoperators on Hγ , respectively. Moreover,

‖φ(x)γx‖H = |φ(x)|√

Γ(x, x) ∈ L1(X,µ)

‖〈·, γx〉Hγx‖tr = Γ(x, x) ∈ L1(X,µ),

where ‖·‖tr is the trace norm. So the integrals are well defined and the claimsconcerning convergence follow.The above result can be restated in the following way. The function γ is(strongly) square-integrable, that is, Eq. (19) holds, if and only if A∗

γAγ is atrace class operator on H and, by polar decomposition, this latter conditionis equivalent to the fact that AγA

∗γ is of trace class on L2(X,µ). This proves

the last statement of the corollary.

The following example shows that the weakly square-integrability is not suf-ficient for compactness of Aγ and that strong square-integrability is not nec-essary.

Example 2 Let H be an infinite dimensional separable Hilbert space and µa measure on X such that there exists a disjoint sequence (Xn)n∈N of mea-surable sets satisfying

0 < µ(Xn) = an < +∞ n ∈ N.

Let (en)n∈N be a basis of H and (σn)n∈N be a positive sequence of R.Define the map γ from X to H as

γx =1√an

σnen if x ∈ Xn.

18

Clearly γ is measurable and, for all v ∈ H,

〈v, γx〉H =1√anσn〈v, en〉H if x ∈ Xn.

It follows that∫

X

|〈v, γx〉H|2dµ(x) =∑

n

σ2n|〈v, en〉H|2.

Since the sequence (|〈v, en〉H|2)n∈N is in ℓ1, γ is a weakly square-integrablefunction if and only if the sequence (σn)n∈N is bounded.Assume from now on that (σn)n∈N is bounded. It follows that, given v ∈ H,

Aγv =∑

n

σn〈v, en〉Hφn

where

φn(x) =

1√an

x ∈ Xn

0 x 6∈ Xn

and the series converges in L2(X,µ).Since (φn)n∈N is an orthonormal sequence of L2(X,µ), the above equationshows that (en, σ

2n) is the spectral decomposition of A∗

γAγ. It follows that Aγ

is a compact operator if and only if

limn→+∞

σn = 0.

Finally, Aγ is a Hilbert-Schmidt operator if and only if

+∞∑

n=1

σ2n < +∞.

We end the section with a simple proposition showing the connection betweenintegral operators and weakly integrable functions (see Halmos and Sunder(1978) for a discussion).

Proposition 7 With the notation of Proposition 3, let K : X × Y → C bea Carleman (p, 2)-bounded kernel with 1 ≤ p ≤ ∞ and let γ : X → L2(Y, ν)

γx = K(x, ·)

for µ-almost all x ∈ X, then γ is a weakly p-integrable function.

19

Proof. By separability of L2(Y, ν), the map x 7→ γx is measurable. Bydefinition of Carleman (p, 2)-bounded kernel, for µ-almost all x ∈ X K(x, ·)is square-integrable with respect to ν and

〈v, γx〉L2(Y ) =

∫

Y

K(x, y)v(y) dν(y) = (LKv)(x) ∀v ∈ L2(Y, ν).

Since K is a (p, 2)-bounded kernel, the range of LK is in Lp(X,µ), so that γis a weakly p-integrable function.

5 Continuity

In this section we study the weak continuity of γ. The following result is dueto Schwartz (1964, Proposition 24), but our proof is based on elementarytools.

Proposition 8 Let X be a locally compact space and H a Hilbert space. Thefollowing facts are equivalent:

1. the map γ is weakly continuous;

2. the function Γ is locally bounded and separately continuous.

If one of the above conditions holds, Im Aγ ⊂ C(X), the operator Aγ iscontinuous from H into C(X), and

ker Aγ = span γx ∈ H|x ∈ X⊥ = Hγ⊥. (20)

If X is separable, then Hγ is separable.

Proof.

1) ⇒ 2) By assumption, Aγv is a continuous function for all v ∈ H. Let x ∈ X,since Γ(·, x) = Aγ(γx)(·), clearly Γ is separately continuous. Fix now acompact set C. Since Aγv = 〈v, γ(·)〉H is continuous, it is bounded onC for all v ∈ H, so the Banach-Steinhaus theorem ensures

‖γx‖H ≤M ∀x ∈ C,

for some constant M > 0. Moreover, by Cauchy-Schwartz inequality,

|Γ(x, t)| = |〈γt, γx〉H| ≤ ‖γx‖H‖γt‖H ≤M2,

for all t, x ∈ C, so that Γ is locally bounded.

20

2) ⇒ 1) Let w =∑n

i=1 aiγxi. The fact that Γ is separately continuous and

Eq. (3) imply that the function Aγw =∑n

i=1 aiΓ(·, xi) is continuous.Let now v ∈ Hγ and x0 ∈ X, we prove that Aγv is continuous at x0.Let C be a compact neighborhood C of x0 and

M = supx∈C

‖γx‖H = supx∈C

√

Γ(x, x),

where M is finite due to locally boundedness of Γ. Fixed ǫ > 0, thereis a finite linear combination w =

∑

aiγxisuch that ‖v − w‖H ≤ ǫ.

By the above observation, Aγw is continuous, so, possibly replacing Cwith a smaller neighborhood, |(Aγw)(x)−(Aγw)(x0)| ≤ ǫ for all x ∈ C.Then

|(Aγv)(x) − (Aγv)(x0)| ≤ |(Aγw)(x) − (Aγw)(x0)| + ‖γx − γx0‖H‖w − v‖H

≤ ǫ(1 + 2M) ∀x ∈ C.

Finally, if v ∈ Hγ⊥, Aγv = 0, so that γ is weakly continuous.

If one of the above equivalent conditions holds, Aγ is a bounded operatorfrom H into C(X). Indeed, let C be a compact set, by locally boundednessof Γ and ‖γx‖2 = Γ(x, x), there is a constant M > 0 such that ‖γx‖H ≤ M

for all x ∈ C. Finally,

supx∈C

|(Aγv)(x)| = supx∈C

|〈v, γx〉H| ≤ ‖v‖H supx∈C

‖γx‖H ≤M‖v‖H,

so that Aγ is bounded. The fact that ker Aγ = Hγ⊥ is clear. Finally, assume

that X is separable, we prove that Hγ is separable. Indeed, let X0 be a densedenumerable subset of X and define

H0 = spanγx | x ∈ X0 ⊂ Hγ .

Clearly, H0 is separable. We claim that H0 = Hγ . Choose v ∈ H⊥0 so that

(Aγv) (x) = 〈v, γx〉H = 0 for all x ∈ X0. Since Aγv is continuous and X0 isdense, 〈v, γx〉H = 0 for all x ∈ X so that v ∈ Hγ

⊥. It follows that H⊥0 ⊂ Hγ

⊥

so that Hγ ⊂ H0 and the claim follows.

The following corollary characterizes the strong continuity of γ (see Propo-sition 24 of Schwartz (1964)).

Corollary 2 With the assumptions of Proposition 8, the following facts areequivalent:

1. the function Γ is continuous;

21

2. the function Γ is continuous on the diagonal of X ×X;

3. the map γ is continuous;

4. the map γ is weakly continuous and the restriction of Γ on the diagonalx 7→ Γ(x, x) is continuous;

5. the operator Aγ : H → C(X) is compact.

Proof.

1) ⇒ 2) Trivial.

2) ⇒ 3) It follows observing that Eq. (3) gives

‖γx − γt‖H2 = Γ(x, x) + Γ(t, t) − Γ(x, t) − Γ(t, x) x, t ∈ X.

3) ⇒ 1) Use the fact that the scalar product is continuous and Eq. (3).

3) ⇐⇒ 4) It is a restatement of the equivalence between strong convergence andweak convergence plus convergence of norms (so called H-property).

3) ⇐⇒ 5) Let B1 be the closed unit ball in H. The operator Aγ is compact if andonly if Aγ(B1) is compact in C(X). Since a locally compact Hausdorffspace is regular, by Ascoli theorem (Kelley, 1955) the set of functionsAγ(B1) is compact if and only if Aγ(B1) is closed, the set (Aγ(B1))(x)is bounded for all x ∈ X, and Aγ(B1) is equicontinuous. First twoconditions are always satisfied since Aγ is continuous from the Hilbertspace H to C(X). Moreover, given x, t ∈ X, we have that

‖γx − γt‖H = supv∈B1

|〈v, γx − γt〉H|

= supv∈B1

|(Aγv)(x) − (Aγv)(t)|

= supf∈Aγ(B1)

|f(x) − f(t)|. (21)

Eq. (21) shows that Aγ(B1) is equicontinuous if and only if map γ iscontinuous.

Let now X be a locally compact second countable Hausdorff space endowedwith a positive Radon measure µ. Assume that γ is a weakly continuous mapfrom X into a Hilbert space H and regard Aγ as an operator from H intoL0(X,µ). Since Hγ is separable, Proposition 2 holds, so that kerAγ = Hµ

⊥.

22

On the other hand, since f ∈ C(X) is equal to 0 in L0(X,µ) if and only iff(x) = 0 for all x ∈ supp µ,

ker Aγ = v ∈ H|〈v, γx〉H = 0 ∀x ∈ supp µ = span γx|x ∈ supp µ⊥,

henceHµ = span γx | x ∈ supp µ. (22)

Finally, assume X compact, so that µ is finite. Since γ is weakly continuous,it is bounded and, hence, strongly p-integrable, so the operator Aγ is alwayscompact as a map in Lp(X,µ). However, in order Aγ be compact as a mapin C(X), it is necessary (and sufficient) that γ is strongly continuous.

6 Mercer theorem

In this section we characterize those reproducing kernel Hilbert spaces thatare subspaces of L2(X,µ) in terms of the spectral decomposition of theintegral operator LΓ. We recall the definition of complex Radon measure(Dieudonne, 1968). Let R > 0 and Cc((0, R]) the space of compactly sup-ported functions on the interval (0, R], a complex Radon measure on (0, R]is a complex linear form on Cc((0, R]) such that its restriction to Cc([a,R])is bounded for all 0 < a < R. If ρ is a complex Radon measure, there is aunique positive Radon measure |ρ| and a complex measurable function h on(0, R], such that

ρ(φ) =

∫

(0,R]

φ(λ)h(λ)d|ρ|(λ) φ ∈ Cc((0, R]).

and |h(λ)| = 1 for all λ ∈ (0, R]. If φ ∈ L1((0, R], |ρ|), the integral withrespect to ρ is defined by

∫

(0,R]

φ(λ)dρ(λ) =

∫

(0,R]

φ(λ)h(λ)d|ρ|(λ).

The measure |ρ| is called the absolute value of ρ.

Proposition 9 Let X be a locally compact second countable Hausdorff spaceendowed with a positive Radon measure µ such that supp µ = X. Let H be areproducing kernel Hilbert space with kernel Γ. The following conditions areequivalent.

1. H is a subspace of L2(X,µ) ∩ C(X);

2. the kernel Γ is locally bounded, separately continuous, and (2, 2)-bounded.

23

If one of the above assumptions holds, the integral operator LΓ of kernel Γ isa positive operator with spectral decomposition

LΓ =

∫

[0,R]

λdP (λ) R = ‖LΓ‖

and

H = v ∈ L2(X,µ) ∩ C(X) |∫

[0,R]

1

λ〈dP (λ)v, v〉L2(X) < +∞ (23)

‖v‖H2 =

∫

[0,R]

1

λ〈dP (λ)v, v〉L2(X) ∀v ∈ H. (24)

If x, y ∈ X there is a complex Radon measure ρx,y on (0, R] such that

Γ(x, y) =

∫

(0,R]

λdρx,y(λ) (25)

and, for all E ∈ B((0, R]) such that E ⊂ (0, R], then

∫

E

dρx,y(λ) =∑

n∈I

φn(x)φn(y) (26)

where φn ∈ H and (φn)n∈I is an orthonormal basis in L2(X,µ) for the rangeof P (E).If x ∈ X, ρx,x is a positive Radon measure on (0, R] such that

ρx,x((0, R]) =∑

n∈I

|φn(x)|2, (27)

where φn ∈ H and (φn)n∈I a basis in L2(X,µ) for ker L⊥γ . In particular ρx,x

is finite if and only if Γ(·, x) ∈ Im LΓ.

Proof. Proposition 1 ensures that the inclusion of H into CX is the operatorAγ associated with the map γ : X → H

γx = Γ(·, x) x ∈ X.

Proposition 8 imply that H ⊂ C(X) if and only if Γ is separately continuousand locally bounded. In particular, since X is separable, both the aboveconditions and Eq. (11) ensure that H is separable. So Proposition 5 withp = 2 gives that H ⊂ L2(X,µ) if and only if Γ is a (2, 2)-bounded kernel.Hence the equivalence of Condition 1 and Condition 2 is now clear. Assumeone of them and regard the inclusion Aγ as an operator from H into L2(X,µ).

24

By Eqs. (22), (13), Aγ is injective and the polar decomposition of the adjointA∗

γ gives

A∗γ = W (AγA

∗γ)

1

2 (28)

where W is a partial isometry from L2(X,µ) to H with kernel being equalto the kernel of A∗

γ and with range being equal to the closure of the range ofA∗

γ .Proposition 5 states that AγA

∗γ = LΓ, so that LΓ is a positive operator,

the kernel of W is the kernel of LΓ and, since Aγ is injective, W is surjective,that is, WW ∗ is the identity. Equation (28) gives

Aγ = L1

2

ΓW∗. (29)

Since Aγ is the inclusion of H into L2(X,µ)∩C(X) and is injective, Eq. (29)

implies that we can identify H with Im L1

2

Γ and, by means of spectral theorem,Eq. (23) follows. In particular, for all v ∈ H, v ∈ (ker LΓ)⊥ = (Im P (0))⊥.

Let now v ∈ H, then v = L1

2

ΓW∗v and

∫

[0,R]

1

λ〈dP (λ)v, v〉L2(X) =

∫

[0,R]

1

λ

⟨

dP (λ)L1

2

ΓW∗v, L

1

2

ΓW∗v

⟩

L2(X)

=

∫

[0,R]

〈dP (λ)W ∗v,W ∗v〉L2(X)

= 〈W ∗v,W ∗v〉L2(X)

= 〈v, v〉H,where we used that P (0)W ∗ = 0. So Eq. (24) follows.

Given x, y ∈ X, let ρx,y be the linear form on Cc((0, R]) given by

ρx,y(φ) =

∫

(0,R]

1

λφ(λ) 〈dP (λ)W ∗γy,W

∗γx〉L2(X) .

Since 〈dP (λ)W ∗γy,W∗γx〉 is a bounded complex measure and φ has com-

pact support, ρx,y is well defined and is a complex Radon measure. Bydefinition, |ρx,y| has density 1

λwith respect to the positive bounded mea-

sure | 〈dP (λ)W ∗γy,W∗γx〉 |. Since the function 1 is integrable with respect

to | 〈dP (λ)W ∗γy,W∗γx〉 |, the function λ is integrable with respect to |ρx,y|

and, hence, with respect to ρx,y. In particular,∫

(0,R]

λdρx,y =

∫

(0,R]

〈dP (λ)W ∗γy,W∗γx〉L2(X)

=

∫

[0,R]

〈dP (λ)W ∗γy,W∗γx〉L2(X)

= 〈W ∗γy,W∗γx〉L2(X) = 〈γy, γx〉H,

25

so that Eq. (25) follows by definition of Γ.Let now E ∈ B((0, R]) be such that E ⊂ (0, R]. This last fact and

the spectral theorem imply that Im P (E) ⊂ Im L1

2

Γ = H. Hence, there is asequence (φn)n∈I of H such that (φn)n∈I is an orthonormal basis for Im P (E).Since E ⊂ (0, R], χE is integrable with respect to ρx,y and

∫

(0,R]

χE(λ)dρx,y =

∫

[0,R]

χE(λ)

λ〈dP (λ)W ∗γy,W

∗γx〉L2(X)

=

∫

[0,R]

1

λ〈dP (λ)P (E)W ∗γy, P (E)W ∗γx〉L2(X)

=⟨

L− 1

2

Γ P (E)W ∗γy, L− 1

2

Γ P (E)W ∗γx

⟩

L2(X), (30)

where P (E)W ∗γx and P (E)W ∗γy are in Im L1

2

Γ .Let now J a finite subset of I, taking into account the properties of the

sequence (φn)n∈I ,

∑

n∈J

⟨

L− 1

2

Γ P (E)W ∗γy, φn

⟩

L2(X)

⟨

φn, L− 1

2

Γ P (E)W ∗γx

⟩

L2(X)

=∑

n∈J

⟨

W ∗γy, L− 1

2

Γ φn

⟩

L2(X)

⟨

L− 1

2

Γ φn,W∗γx

⟩

L2(X)

(φn = L1

2

ΓW∗φn)

=∑

n∈J

〈W ∗γy,W∗φn〉L2(X) 〈W ∗φn,W

∗γx〉L2(X)

=∑

n∈J

〈γy, φn〉H〈φn, γx〉H

(Eq. (8))

=∑

n∈J

φn(x)φn(y).

Eq. (26) follows observing that the series

∑

n∈I

⟨

L− 1

2

Γ P (E)W ∗γy, φn

⟩

L2(X)

⟨

φn, L− 1

2

Γ P (E)W ∗γx

⟩

L2(X)

is summable with sum⟨

L− 1

2

Γ P (E)W ∗γy, L− 1

2

Γ P (E)W ∗γx

⟩

L2(X)= ρx,y(E), by

means of Eq. (30).Finally, if x ∈ X, clearly ρx,x is a positive Radon measure on (0, R] having

density 1λ

with respect to the positive bounded measure 〈dP (λ)W ∗γx,W∗γx〉.

26

Hence ρx,x is bounded if and only if W ∗γx ∈ Im L1

2

Γ . Equation (29) impliesthat this condition is equivalent to γx ∈ Im LΓ and, if it is satisfied,

ρx,x((0, R]) =∥

∥

∥L− 1

2

Γ W ∗γx

∥

∥

∥

2

L2(X). (31)

Let (φn)n∈N be a sequence in H such that (φn)n∈N is a basis in L2(X,µ) ofker L⊥

Γ . Given N ∈ N, reasoning as above,

N∑

n=1

|φn(x)|2 =N

∑

n=1

∣

∣

∣

∣

⟨

L− 1

2

Γ φn,W∗γx

⟩

L2(X)

∣

∣

∣

∣

2

.

The series in the right side converges if and only if W ∗γx ∈ Im L1

2

Γ and, if it

is so, its sum is∥

∥

∥L− 1

2

Γ W ∗γx

∥

∥

∥

2

L2(X). Eq. (31) implies Eq. (27).

Equation (23) allows us to identify the elements of H with the only continuous

functions on X whose equivalence class belongs to the range of L1

2

Γ , extendinga result of Cucker and Smale (2002). The assumptions that supp µ = X andH ⊂ C(X) ensure that the identification between functions in H and equiv-alence classes in L2(X,µ) is well defined. With this identification, Eq. (24)

implies that L1

2

Γ is a unitary operator from ker L⊥Γ onto H.

If supp µ 6= X, the statements of the above proposition hold replacing X

with supp µ and H with span γx|x ∈ supp µ. We use the assumption thatγ is weakly continuous only in two steps: to show the separability of H and toidentify elements of H, which are functions on X, with elements in L2(X,µ),which are equivalence classes.

If the integral operator LΓ has a pure point spectrum

LΓ =∑

n∈I

λn 〈·, φn〉φn,

where λn ≥ 0 and (φn)n∈N is a basis of L2(X,µ), it is possible to chooseφn ∈ H for all λn > 0 and, with this choice, Proposition 9 gives that

H = v ∈ L2(X,µ) ∩ C(X) |∑

n

1

λn| 〈v, φn〉L2(X) |2 < +∞,

‖v‖2H =

∑

n

1

λn

| 〈v, φn〉L2(X) |2,

Γ(x, y) =∑

n

λnφn(x)φn(y).

The last series converges absolutely and, by Dini theorem and Condition 4of Corollary 2, converges uniformly on compact subsets if and only if γ isstrongly continuous.

27

References

Syed Twareque Ali, Jean-Pierre Antoine, and Jean-Pierre Gazeau. Coherentstates, wavelets and their generalizations. Graduate Texts in Contempo-rary Physics. Springer-Verlag, New York, 2000. ISBN 0-387-98908-0.

N. Aronszajn. Theory of reproducing kernels. Trans. Amer. Math. Soc., 68:337–404, 1950.

M. Bachir Bekka and Pierre de la Harpe. Irreducibility of unitary group rep-resentations and reproducing kernels Hilbert spaces. Expo. Math., 21(2):115–149, 2003. ISSN 0723-0869. Appendix by the authors in collaborationwith Rostislav Grigorchuk.

Nicolas Bourbaki. Integration. I. Chapters 1–6. Elements of Mathematics(Berlin). Springer-Verlag, Berlin, 2004. ISBN 3-540-41129-1. Translatedfrom the 1959, 1965 and 1967 French originals by Sterling K. Berberian.

John B. Conway. A course in functional analysis, volume 96 of GraduateTexts in Mathematics. Springer-Verlag, New York, second edition, 1990.ISBN 0-387-97245-5.

Felipe Cucker and Steve Smale. On the mathematical foundations of learning.Bull. Amer. Math. Soc. (N.S.), 39(1):1–49 (electronic), 2002. ISSN 0273-0979.

Ingrid Daubechies. Ten lectures on wavelets, volume 61 of CBMS-NSF Re-gional Conference Series in Applied Mathematics. Society for Industrialand Applied Mathematics (SIAM), Philadelphia, PA, 1992. ISBN 0-89871-274-2.

J. Dieudonne. Elements d’analyse. Tome II: Chapitres XII a XV. CahiersScientifiques, Fasc. XXXI. Gauthier-Villars, Editeur, Paris, 1968.

Dorin Ervin Dutkay. Positive definite maps, representations and frames. Rev.Math. Phys., 16(4):451–477, 2004. ISSN 0129-055X.

Roger Godement. Les fonctions de type positif et la theorie des groupes.Trans. Amer. Math. Soc., 63:1–84, 1948.

Paul Richard Halmos and Viakalathur Shankar Sunder. Bounded integraloperators on L2 spaces, volume 96 of Ergebnisse der Mathematik und ihrerGrenzgebiete [Results in Mathematics and Related Areas]. Springer-Verlag,Berlin, 1978. ISBN 3-540-08894-6.

28

Einar Hille. Introduction to general theory of reproducing kernels. RockyMountain J. Math., 2(3):321–368, 1972.

Einar Hille and Ralph S. Phillips. Functional analysis and semi-groups.American Mathematical Society, Providence, R. I., 1974. Third printingof the revised edition of 1957, American Mathematical Society ColloquiumPublications, Vol. XXXI.

Harry Hochstadt. Integral equations. Wiley Classics Library. John Wiley& Sons Inc., New York, 1989. ISBN 0-471-50404-1. Reprint of the 1973original, A Wiley-Interscience Publication.

John L. Kelley. General topology. D. Van Nostrand Company, Inc., Toronto-New York-London, 1955.

A.N. Kolmogorov. Stationary sequences in Hilbert space. In Selected works.Probability theory and mathematical statistics, volume II, pages 228–271.Kluwer, 1992.

M. G. Kreın. Hermitian positive kernels on homogeneous spaces. I. Ukrain.Mat. Zurnal, 1(4):64–98, 1949.

M. G. Kreın. Hermitian-positive kernels in homogeneous spaces. II. Ukrain.Nat. Zurnal, 2:10–59, 1950.

I.M. Novitskii and M.A. Romanov. An extension of Mercer’s theorem tounbounded operator. Far Eastern Mathematical Reports, 7:123–132, 1999.(in Russian, English translation: arXiv:math.FA/0303159v3).

B. J. Pettis. On integration in vector spaces. Trans. Amer. Math. Soc., 44(2):277–304, 1938. ISSN 0002-9947.

S. Saitoh. Integral transforms, reproducing kernels and their applications,volume 369 of Pitman Research Notes in Mathematics Series. Longman,Harlow, 1997. ISBN 0-582-31758-4.

Saburou Saitoh. Theory of reproducing kernels and its applications, volume189 of Pitman Research Notes in Mathematics Series. Longman Scientific& Technical, Harlow, 1988. ISBN 0-582-03564-3.

I. J. Schoenberg. Metric spaces and positive definite functions. Trans. Amer.Math. Soc., 44(3):522–536, 1938. ISSN 0002-9947.

29

Laurent Schwartz. Sous-espaces hilbertiens d’espaces vectoriels topologiqueset noyaux associes (noyaux reproduisants). J. Analyse Math., 13:115–256,1964.

Laurent Schwartz. Analyse. III, volume 44 of Collection Enseignement desSciences [Collection: The Teaching of Science]. Hermann, Paris, 1993.ISBN 2-7056-6163-2. Calcul integral.

Robert M. Young. An introduction to nonharmonic Fourier series. AcademicPress Inc., San Diego, CA, first edition, 2001. ISBN 0-12-772955-0.

30